Plot results for CIFAR-10 data

from pathlib import Path

from src.model import load_histories
from src.plots import plot_histories

path = Path().resolve() / 'experiments' / 'cifar10'

The plots always show three plots for loss and accuracy:

• The drop in training and test data (i.e. the difference of the second and third plot - to illustrate the extend of overfitting)
• The value based on training data
• The value based on test data

The legend means:

l_"Type of label" _ s _ "strength of regularization" _ lr _ "learning rate"

s = 0.0 means no regularization, s= 1.0 is the standard regularization. s=5.0 and s=10.0 are stronger regularizations. The threshold for the evidence regularization is always 10% of the batch_size, i.e. 0.1 * 1024 = 102.4

Clean Label (the normal label)

patterns = ['l_clean_label', 'lr_0.001']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_clean_label', 'lr_0.01']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_clean_label', 'lr_0.1']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_clean_label', ['s_0.0', 's_1.0', 's_10.0'], 'lr_1.0']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_clean_label']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

## Noisy label

patterns = ['l_aggre_label', 'lr_0.001']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_aggre_label', 'lr_0.01']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_aggre_label', 'lr_0.1']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_aggre_label', 'lr_1.0']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_aggre_label']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

## Noisy label (worse) -> more noise

patterns = ['l_worse_label', 'lr_0.001']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_worse_label', 'lr_0.01']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_worse_label', 'lr_0.1']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_worse_label', 'lr_1.0']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)

patterns = ['l_worse_label']
histories = load_histories(path, patterns)
fig, ax = plot_histories(histories)